Geek alert (headline = link btw). That is an excellent piece on overfitting effects in what many people refer to as ‘AI’ (Artificial Intelligence) or ‘ML’ (Machine Learning).
Overfitting: the nature of the models widely used, in detail some methods like backpropagation, amplifies multicollinearity within these models, which basically means to artificially create stronger correlations between independent variables and the dependent variable in a development sample. That can lead to extraordinary predictive accuracy of the model – which is limited to the test sample, while it performs much worse using e.g. real-life data. The model then has been ‘over-fitted’ to that particular data sample.
Despite (European) startups claiming to use ‘AI’ in many cases but just don’t do that, always remember as a rule of thumb if someone talks to you about ‘AI’:
In your pitch, it’s AI.
In your specs, it’s ML.
In your pilot, it’s linear regression.
In your product, it’s printf();
Don’t know the source (thanks for any hint!), I recently read it on LinkedIn and had a good laugh 🙂